CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-09 10:17:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama.cpp : use rope scale kv

Browse Source
This commit is contained in:
klosax
2023-08-21 13:30:03 +02:00
committed by GitHub
parent dc1f051013
commit c082b9fa0b
1 changed files with 8 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
llama.cpp
View File

@@ -1435,6 +1435,14 @@ static void llama_model_load_internal(
hparams.n_head_kv = hparams.n_head; hparams.n_head_kv = hparams.n_head;
GGUF_GET(hparams.n_head_kv, gguf_get_val_u32, GGUF_TYPE_UINT32, false, "llama.attention.head_count_kv"); GGUF_GET(hparams.n_head_kv, gguf_get_val_u32, GGUF_TYPE_UINT32, false, "llama.attention.head_count_kv");
// TODO: manually setting rope scale should override this
// rope_freq_scale (inverse of the kv) is optional
float ropescale = 1.0f;
GGUF_GET(ropescale, gguf_get_val_f32, GGUF_TYPE_FLOAT32, false, "llama.rope.scale_linear");
if (ropescale != 1.0f) {
rope_freq_scale = 1.0f/ropescale;
}
// get general kv // get general kv
GGUF_GET(general_name, gguf_get_val_str, GGUF_TYPE_STRING, false, "general.name"); GGUF_GET(general_name, gguf_get_val_str, GGUF_TYPE_STRING, false, "general.name");
GGUF_GET(general_arch, gguf_get_val_str, GGUF_TYPE_STRING, false, "general.architecture"); GGUF_GET(general_arch, gguf_get_val_str, GGUF_TYPE_STRING, false, "general.architecture");